Announcing Mistral Large 4: Le Chonk
Mistral Large 4 preview: 1T total, 49B active parameters, API-only access, two reasoning levels, Artificial Analysis 38, and open weights soon.
Key Takeaways
- Mistral Large 4 preview, nicknamed “Le Chonk,” is available only through Mistral’s API today, with 1 trillion total parameters and 49 billion active parameters.
- The model exposes two reasoning levels, “none” and “high,” and the “high” response appears stronger even though it used fewer output tokens in the sample comparison.
- On Artificial Analysis, the preview scores 38, trailing DeepSeek 4.1 Flash but marking a major improvement over Mistral Large 3, which scored 9.
- It is not a frontier model, being described as roughly six months behind current leaders, but may be useful for teams seeking large Mistral API access and future open-weight potential.
- The key next milestone is the promised open-weights release by “end of this month,” which would enable local deployment, fine-tuning and evaluation, though exact license, hardware needs and final performance remain unclear.
Headline: Mistral Large 4 Preview Brings 1 Trillion Parameters and API Access
Meta description: Mistral has released a preview of Mistral Large 4, a 1 trillion parameter model with 49 billion active parameters, available via its API today.
Tags: ai, generative-ai, llms, mistral, llm-release
Featured image: An abstract, friendly illustration of a large model preview, with a stylized pelican shape and API icons, no real people, no logos, no text
What the preview offers today
Mistral Large 4 preview arrived through Mistral's API today, giving developers, researchers and product teams early access to a model the company describes as having 1 trillion total parameters and 49 billion active parameters. The release is being called "Le chonk" in the original announcement, and it is notable because Mistral is trying to stay visible in a crowded model race while promising an open weights model at "end of this month." For now, the practical path is hosted: users can send requests to the API, compare outputs and plan evaluation, but they cannot download the weights yet. The source does not give a more precise calendar date for the open weights release, nor does it describe pricing, rate limits or a full evaluation suite.
Why the parameter split matters
The headline size is 1 trillion parameters, but the 49 billion active parameter figure is the one that affects day-to-day inference. In a large model, not every parameter is necessarily used for every token; the active subset controls how much computation is spent during generation. That matters for teams comparing API performance, latency and cost, and for later open-weight deployments where hardware availability is a central concern. Mistral says the model was trained on its own cluster of 3,800 NVIDIA Grace Blackwell GPUs. The cluster detail suggests a substantial internal training setup, though the source does not explain the full training recipe, data mix or exact deployment requirements.
Reasoning levels and the pelican comparison
The API currently exposes only two reasoning levels: "none" and "high." That is a simpler control surface than some models, and it gives early testers a clear way to compare the model's visible reasoning behavior. The source shows two pelican outputs, tied to the pelican-riding-a-bicycle tag, and the "high" result looks better than the "none" result. The token counts are counterintuitive: the "high" example used 2,717 output tokens, while the "none" example used 3,275. In other words, the more impressive-looking response did not come with a larger token count, which may matter to developers who assume reasoning quality always increases cost.
How it compares on Artificial Analysis
On Artificial Analysis, a model comparison site referenced in the source, the preview scores 38. The source places it just behind DeepSeek 4.1 Flash, which is described as a 552B model. That comparison is important because it shows the release is not only a size story; it is being judged against another large model from a different lab. The larger context is Mistral's own trajectory. Last December's Mistral Large 3 received a much weaker result on the same Artificial Analysis view, scoring 9, and the source describes its pelican output as terrible. The new score is therefore a major improvement for Mistral, even if it does not put the model at the very top of the leaderboard.
The source is explicit about the limits: the model is not a Fable-class model. Still, the assessment is that Mistral has put out something that is maybe about 6 months behind the frontier. For readers evaluating whether to test it, that gap is meaningful. It may not be the first choice for a product that needs the absolute latest capabilities, but it could be useful for teams that want a large Mistral model with open-weight momentum, API access today and a clearer comparison with DeepSeek 4.1 Flash and Mistral Large 3.
What developers should watch next
The next key fact is whether the open weights model actually appears by "end of this month." If it does, the value for researchers and enterprises could change quickly: downloadable weights allow local experimentation, custom fine-tuning, auditing and deployment outside Mistral's hosted API. The current preview is useful for early evaluation, but it leaves questions about the final license, exact hardware needs and how the open version will compare with the API version.
For developers, the immediate actions are simple. Test both "none" and "high" reasoning levels on a representative workload, not just the pelican example, and track output tokens, latency and task quality. If the open weights release is important, treat the "end of this month" promise as a planning date, not a guarantee, because the source does not provide a more exact day. The Hacker News-sourced report frames this as a comeback moment for Mistral, but the longer-term signal will be whether the API preview becomes a practical, open-weight alternative in a market where model size, active parameter efficiency, benchmark scores and release cadence all influence adoption.

Comments (0)